Bayesian Enhanced α-Expansion Move Clustering with Loose Link Constraints

Abstract

Pairwise link constraints, as an auxiliary information, can help improve the clustering performances a lot. Yet, among them loose link constraints can be acquired more easily and cheaply and hence are more widely utilized in practical applications compared with strong link constraints. Therefore, in this paper, we focus on exemplar-based clustering with loose link constraints. Based on Bayesian probabilistic framework, we naturally integrate the Enhanced α–Expansion Move (EEM) clustering algorithm with loose link constraints, and accordingly propose the Bayesian Enhanced α-Expansion Move Clustering (BEEMLC) algorithm with Loose Link Constraints. The proposed clustering algorithm BEEMLC can exhibit the very applicability of the enhanced α-expansion move clustering in the following two aspects: 1) BEEMLC originates from EEM yet retains the basic spirit of the optimization algorithm contained in EEM. In fact, we directly add a penalty term about loose link constraints into the objective function. Therefore it indeed inherits the advantages of EEM in improving clustering performance but extends such advantages into clustering with loose link constraints. 2) In contrast to other semi-supervised Affinity Propagation clustering algorithms, BEEMLC indeed deals with loose link constraints rather than strong link constraints only. Experiments on benchmarking and real-world datasets, as well as the application of user interactive image segmentation, have shown comparable and even better performance of BEEMLC, compared with other state-of-the-art exemplar-based clustering algorithms.